Security

We said: rules before code. Here's the code.

A July 2026 article about the first version of Geslar CLI and the five-rule doctrine, updated for CLI 1.0. The current state is described at docs.geslar.app.

updated 4 min read

Contents

The problem we're solving

A few weeks ago we published a five-rule doctrine for how AI tools should access passwords — before we'd written a single line of code. We said rules come before code, so feel free to hold us to it.

In July 2026 the first version of the code shipped. Note (updated for the 1.0 release): the code in this article describes that first version, and commands and options have changed since (for example, --env-file is now -f); the current ones are at docs.geslar.app.

AI agents — Claude Code, Cursor and similar tools — increasingly run our apps, tests and deploys. To do that, they need access to secrets: database passwords, API keys, tokens. The usual way this gets solved today is simply handing the secret to the agent. Into its context. Where the model reads it, where it stays in the conversation history, and where a prompt injection attack can extract it — and no reliable defence against that exists yet.

We started from the opposite assumption: a secret must never enter the AI context. The agent may orchestrate — start, connect, execute — but the secret's value gets injected directly into the program that needs it, never into the conversation with the model.

How Geslar CLI does it

Your .env file no longer holds secrets, just references:

DATABASE_URL=geslar://Work/Database/url
STRIPE_KEY=geslar://Work/Stripe/secret_key
You can safely commit that file — there's nothing secret in it. You run the app like this:

geslar run -f .env -- npm start

Geslar resolves the references and decrypts the values on your machine (the server never sees your keys — the same zero-knowledge principle as the rest of Geslar), and injects them directly into the process you're running. The secrets exist only in that process's memory, while it runs. If any reference is invalid — nothing starts (better to fail loudly than run on half a configuration).
In that version the agent saw only the reference (geslar://…) and the run's result, and the value did not pass through it. With 1.0 that also holds for agents with their own profile: through the MCP server (geslar mcp serve) the model receives the exit code, the duration and the number of output lines, never the value. The first read of a vault's items has to be approved by a person, with a password and a fresh 2FA code (until then nothing runs). A CI job uses a service account and GESLAR_SERVICE_TOKEN, and there is a GitHub Action. The one way output can reach the model is the opt-in --return-stdout-tail, a filtered tail that is best effort and not a security boundary; very short values are not masked in it.

Five rules, now in code

The doctrine is no longer a manifesto — it's how the tool behaves:
A secret never enters the AI context
geslar run injects into the process, not the conversation.
Every access is approved
An agent's first read of a vault's items is approved by a person, with a password and a fresh 2FA code.
Least privilege, limited duration
Unlocking has a deadline (TTL); locking is instant.
Decryption only on your device
Secret values are decrypted on your computer; the server does not see them.

Honest about scope

Note (updated for the 1.0 release): this article was written for the first CLI release, and the CLI has changed since. What it does not protect against: while your vault is unlocked, anything that runs as you — an AI agent with access to your terminal included — can use it without approval. Run geslar lock before you let an agent work in your terminal, or run the agent under a separate OS user. Masking hides only the literal value. The current description and the full list of limitations are at docs.geslar.app/cli/limitations.

Start in a minute

npm install -g @geslar/cli (Geslar CLI is available with a Geslar account; see which package includes what on the pricing page)

geslar mcp init <profile> wires an agent profile into Claude Code, Claude Desktop or Cursor, and any MCP client that can start a local stdio server can use geslar mcp serve. There is no hosted version, by design: your secrets are decrypted only on your own computer.

Full documentation and quickstart are at docs.geslar.app.
We're building Geslar so an AI agent can work for you — but must never see your passwords. If you're letting AI tools near your credentials, give it a try and tell us where it still hurts.

How we solve this at Geslar

The .env file holds no secrets, only references of the form geslar://, and the values are decrypted solely on your own computer. A person approves an agent's first read of vault entries.

Read more about security →
Make a password an attacker will not guessA generator of Croatian phrases — free, no account.
Try the generator →
Author
Geslar
Geslar d.o.o. builds Škrinjar and Vezar — password and access protection for people and businesses.
More about Geslar →